In [6]:
y = y.tail(n=200) # approximately 4 years of data
y.plot(ylabel='Weekly Series '+series_id, grid=True)
Out[6]:
<Axes: ylabel='Weekly Series W142'>
by the Oracle AutoMLx Team
Forecasting Demo notebook.
Copyright © 2024, Oracle and/or its affiliates.
Licensed under the Universal Permissive License v 1.0 as shown at https://oss.oracle.com/licenses/upl/
In this notebook we will build a forecaster using the Oracle AutoMLx tool for three real-world datasets. We explore the various options available in the Oracle AutoMLx Forecasting module, allowing the user to control the AutoML training process. We finally evaluate the forecasting algorithms using in-built visualization tools. Finally, we provide an overview of the capabilities that Oracle AutoMLx offers for explaining the predictions of the tuned model.
Forecasting uses historical time series data as input to make informed estimates of future trends. Learning an accurate forecasting model requires expertise in data science and statistics. This process typically comprises of:
These steps are significantly time consuming and heavily rely on data scientist expertise. Unfortunately, to make this problem harder, the best feature subset, model, and hyperparameter choice widely varies with the dataset and the prediction task. Hence, there is no one-size-fits-all solution to achieve reasonably good model performance. Using a simple Python API, AutoML can quickly jump-start the datascience process with an accurately-tuned model and appropriate features for a given prediction task.
%matplotlib inline
%load_ext autoreload
%autoreload 2
Load the required modules.
import time
import datetime
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
try:
from sktime.forecasting.model_selection import temporal_train_test_split
except ImportError:
try:
from sktime.split import TemporalTrainTestSplitter as temporal_train_test_split
except ImportError as e:
raise ImportError("Failed to import Splitters from sktime. "
"Please ensure you have the correct version of sktime installed.") from e
plt.rcParams['figure.figsize'] = [15, 5]
plt.rcParams['font.size'] = 15
import automlx
from automlx import init
We fetch the series from the repository of the M4 forecasting competition to use throughout this demo.
! wget https://github.com/Mcompetitions/M4-methods/raw/master/Dataset/Train/Weekly-train.csv -q
! wget https://github.com/Mcompetitions/M4-methods/raw/master/Dataset/M4-info.csv -q
all_series = pd.read_csv("Weekly-train.csv", index_col=0) # consists of thousands of series
metadata_csv = pd.read_csv("M4-info.csv", index_col=0) # describes their datetime index
The Oracle AutoMLx solution for forecasting can process both univariate (where only a single time series is available) and multivariate time series (where multiple time series are available). We start by displaying an example of use for univariate time series, and will address multivariate data at the end of this notebook.
We select one time series from the finance sector with weekly collection frequency. M4 dataset requires additional preprocessing to reconstruct the time series.
series_id = 'W142'
series_metadata = metadata_csv.loc[series_id]
series_values = all_series.loc[series_id]
# drop NaNs for the time period where data wasn't recorded
series_values.dropna(inplace=True)
# retrieve starting date of recording and series length to generate the datetime index
start_date = pd.to_datetime(series_metadata.StartingDate)
future_dates = pd.date_range(start=start_date,
periods=len(series_values),
freq='W')
y = pd.DataFrame(series_values.to_numpy(),
index=future_dates,
columns=[(series_metadata.category+"_"+series_id)])
We can now visualize the last 200 weeks of data we have on hand.
y = y.tail(n=200) # approximately 4 years of data
y.plot(ylabel='Weekly Series '+series_id, grid=True)
<Axes: ylabel='Weekly Series W142'>
One must ensure that the data points are in a Pandas DataFrame, sorted in chronological order.
print(y.index)
print("Time Index is", "" if y.index.is_monotonic_increasing else "NOT", "monotonic.")
print("Train datatype", type(y))
DatetimeIndex(['2012-12-09 12:00:00', '2012-12-16 12:00:00',
'2012-12-23 12:00:00', '2012-12-30 12:00:00',
'2013-01-06 12:00:00', '2013-01-13 12:00:00',
'2013-01-20 12:00:00', '2013-01-27 12:00:00',
'2013-02-03 12:00:00', '2013-02-10 12:00:00',
...
'2016-07-31 12:00:00', '2016-08-07 12:00:00',
'2016-08-14 12:00:00', '2016-08-21 12:00:00',
'2016-08-28 12:00:00', '2016-09-04 12:00:00',
'2016-09-11 12:00:00', '2016-09-18 12:00:00',
'2016-09-25 12:00:00', '2016-10-02 12:00:00'],
dtype='datetime64[ns]', length=200, freq='W-SUN')
Time Index is monotonic.
Train datatype <class 'pandas.core.frame.DataFrame'>
As can be seen above, the data contains 100 weekly recorded values over the past 4 years. We will try to predict electricity consumption for the last 0.5 year of data (26 data points), using the previous years as training data. Hence, we separate the dataset into training and testing sets using Temporal train-test split, which ensures that the continuity of the input time series is preserved. Each point in the series represents a month, so we will hold out the last 26 points as test data.
y_train, y_test = temporal_train_test_split(y, test_size=26)
print("Training length: ", len(y_train)," Testing length: ", len(y_test))
Training length: 174 Testing length: 26
print("y_train", y_train)
print("\ny_test", y_test)
y_train Finance_W142 2012-12-09 12:00:00 4793.269 2012-12-16 12:00:00 4818.969 2012-12-23 12:00:00 4863.783 2012-12-30 12:00:00 4926.357 2013-01-06 12:00:00 4916.616 ... ... 2016-03-06 12:00:00 5024.759 2016-03-13 12:00:00 5021.153 2016-03-20 12:00:00 4996.524 2016-03-27 12:00:00 5004.032 2016-04-03 12:00:00 5053.536 [174 rows x 1 columns] y_test Finance_W142 2016-04-10 12:00:00 5039.544 2016-04-17 12:00:00 5053.420 2016-04-24 12:00:00 5066.860 2016-05-01 12:00:00 5099.887 2016-05-08 12:00:00 5082.291 2016-05-15 12:00:00 5084.340 2016-05-22 12:00:00 5081.651 2016-05-29 12:00:00 5106.797 2016-06-05 12:00:00 5147.906 2016-06-12 12:00:00 5140.993 2016-06-19 12:00:00 5153.210 2016-06-26 12:00:00 5258.369 2016-07-03 12:00:00 5307.455 2016-07-10 12:00:00 5295.631 2016-07-17 12:00:00 5300.088 2016-07-24 12:00:00 5295.456 2016-07-31 12:00:00 5333.739 2016-08-07 12:00:00 5338.803 2016-08-14 12:00:00 5339.390 2016-08-21 12:00:00 5345.043 2016-08-28 12:00:00 5369.168 2016-09-04 12:00:00 5401.186 2016-09-11 12:00:00 5395.298 2016-09-18 12:00:00 5386.552 2016-09-25 12:00:00 5377.376 2016-10-02 12:00:00 5406.685
The AutoML pipeline offers the function init, which allows to initialize the parallelization engine.
init(engine='ray')
[2024-08-15 19:16:09,958] [automlx.backend] Overwriting ray session directory to /tmp/xppf3m98/ray, which will be deleted at engine shutdown. If you wish to retain ray logs, provide _temp_dir in ray_setup dict of engine_opts when initializing the AutoMLx engine.
The Oracle AutoMLx solution automatically provides a tuned forecasting pipeline that best models the given training dataset and a prediction task at hand. Here the dataset can be any univariate time-series.
AutoML for Forecasting consists of three main modules:
These pieces are readily combined into a simple AutoML pipeline which automates the entire forecasting process with minimal user input/interaction. One can then evaluate and visualize the forecast produced by the selected model, and optionally the other tuned models.
The AutoML API is quite simple to work with. We first create an instance of the pipeline. Next, the training data is passed to the fit() function which successively executes the previously mentioned modules.
The generated model can then be used for forecasting tasks. By default, we use the negative of symmetric mean absolute percentage error (sMAPE) scoring metric to evaluate the model performance
est1 = automlx.Pipeline(task='forecasting', n_algos_tuned=2)
est1.fit(X=None, y=y_train)
print('Selected model: {}'.format(est1.selected_model_))
print('Selected model params: {}'.format(est1.selected_model_params_))
[2024-08-15 19:16:14,515] [automlx.data_transform] Number of simple differencing orders required: d = 1
[2024-08-15 19:16:14,522] [automlx.data_transform] Seasonal Periodicities; from decomposed/adjusted: [52, 35, 1]
[2024-08-15 19:16:14,766] [automlx.interface] Dataset shape: (174,2)
[2024-08-15 19:16:15,056] [automlx.data_transform] Running preprocessing. Number of features: 3
[2024-08-15 19:16:15,260] [automlx.data_transform] Preprocessing completed. Took 0.204 secs
[2024-08-15 19:16:15,470] [automlx.process] Running Model Generation
[2024-08-15 19:16:15,602] [automlx] Provided model (VARMAXForecaster) is not supported.Supported models are: ['LGBMForecaster', 'ExtraTreesForecaster', 'ExpSmoothForecaster', 'NaiveForecaster', 'SARIMAXForecaster', 'STLwARIMAForecaster', 'ThetaForecaster', 'ProphetForecaster', 'ETSForecaster', 'STLwESForecaster', 'XGBForecaster']
[2024-08-15 19:16:15,603] [automlx] Provided model (DynFactorForecaster) is not supported.Supported models are: ['LGBMForecaster', 'ExtraTreesForecaster', 'ExpSmoothForecaster', 'NaiveForecaster', 'SARIMAXForecaster', 'STLwARIMAForecaster', 'ThetaForecaster', 'ProphetForecaster', 'ETSForecaster', 'STLwESForecaster', 'XGBForecaster']
[2024-08-15 19:16:15,614] [automlx.process] Model Generation completed.
[2024-08-15 19:16:15,702] [automlx.model_selection] Running Model Selection
[2024-08-15 19:16:25,389] [automlx.dataset] Forecast horizon set to 13 for validation sets.
[2024-08-15 19:16:37,131] [automlx.model_selection] Model Selection completed - Took 21.430 sec - Selected models: ['ETSForecaster', 'STLwARIMAForecaster']
[2024-08-15 19:16:37,626] [automlx.trials] Running Model Tuning for ['InputTargetDataTransformer_ETSForecaster']
[2024-08-15 19:16:37,804] [automlx.dataset] Forecast horizon set to 13 for validation sets.
[2024-08-15 19:16:53,762] [automlx.trials] Best parameters for InputTargetDataTransformer_ETSForecaster: {'error': 'add', 'trend': 'mul', 'damped_trend': True, 'seasonal': 'add', 'seasonal_periods': 52, 'initialization_method': 'estimated', 'initial_level': None, 'initial_trend': None, 'initial_seasonal': None, 'bounds': None, 'dates': None, 'freq': None, 'missing': 'none', 'start_params': None, 'maxiter': 1500, 'disp': -1, 'return_params': False}
[2024-08-15 19:16:53,764] [automlx.trials] Model Tuning completed. Took: 16.138 secs
[2024-08-15 19:16:54,267] [automlx.trials] Running Model Tuning for ['InputTargetDataTransformer_STLwARIMAForecaster']
[2024-08-15 19:16:54,474] [automlx.dataset] Forecast horizon set to 13 for validation sets.
[2024-08-15 19:16:59,525] [automlx.trials] Best parameters for InputTargetDataTransformer_STLwARIMAForecaster: {'seasonal_deg': 0, 'trend_deg': 1, 'low_pass_deg': 0, 'period': 52, 'arima_p': 2, 'arima_d': 1, 'arima_q': 2, 'arima_trend': 'n', 'concentrate_scale': True}
[2024-08-15 19:16:59,527] [automlx.trials] Model Tuning completed. Took: 5.259 secs
[2024-08-15 19:17:03,924] [automlx.interface] Re-fitting pipeline
[2024-08-15 19:17:08,635] [automlx.interface] AutoMLx completed.
Selected model: ETSForecaster
Selected model params: {'error': 'add', 'trend': 'mul', 'damped_trend': True, 'seasonal': 'add', 'seasonal_periods': 52, 'initialization_method': 'estimated', 'initial_level': None, 'initial_trend': None, 'initial_seasonal': None, 'bounds': None, 'dates': None, 'freq': None, 'missing': 'none', 'start_params': None, 'maxiter': 1500, 'disp': -1, 'return_params': False}
There are two interfaces that support generating future forecasts using the trained forecasting pipeline.
The preferred function, forecast(), accepts a user-input value for the number of periods to forecast into the future, i.e., relative to the end of the training series. It also accepts a significance level to generate prediction confidence intervals (CIs). When the methods support intervals, confidence intervals at 1-alpha are generated, for example, significance level alpha=0.05 generates 95% confidence intervals.
predictions = est1.forecast(periods=len(y_test), alpha=0.05)
The predict(X) interface supports absolute index-based forecasts, but does not support confidence intervals (CIs). It also downcasts the index to int64.
It should be utilized only when you want to get both in-sample predictions (predictions at timestamps that are part of the training set) and out-of-sample predictions (predictions at timestamps that are not in the training set). The X variable should be an empty dataframe containing only the requested timestamps as index. Here we request 5 in-sample model fit values and 5 out-of-sample forecasts.
future_index = y_train.index[-5:].union(y_test.index[:5])
print(future_index)
DatetimeIndex(['2016-03-06 12:00:00', '2016-03-13 12:00:00',
'2016-03-20 12:00:00', '2016-03-27 12:00:00',
'2016-04-03 12:00:00', '2016-04-10 12:00:00',
'2016-04-17 12:00:00', '2016-04-24 12:00:00',
'2016-05-01 12:00:00', '2016-05-08 12:00:00'],
dtype='datetime64[ns]', freq='W-SUN')
est1.predict(X=pd.DataFrame(index=future_index))
| Finance_W142 | |
|---|---|
| 2016-03-06 12:00:00 | 5015.627932 |
| 2016-03-13 12:00:00 | 5013.775014 |
| 2016-03-20 12:00:00 | 4998.443940 |
| 2016-03-27 12:00:00 | 5013.689868 |
| 2016-04-03 12:00:00 | 5028.477761 |
| 2016-04-10 12:00:00 | 5052.891415 |
| 2016-04-17 12:00:00 | 5066.990374 |
| 2016-04-24 12:00:00 | 5080.142527 |
| 2016-05-01 12:00:00 | 5104.394553 |
| 2016-05-08 12:00:00 | 5090.115109 |
AutoML provides a simple one-line tool to visualize forecasts and confidence intervals.
est1.plot_forecast(predictions=predictions, additional_frames=dict(y_test=y_test))
During the AutoML process, a summary of the optimization process is logged. It consists of:
AutoML provides a print_summary() API to output all the different trials performed.
est1.print_summary()
| Training Dataset size | (174, 2) |
|---|---|
| Validation Dataset size | None |
| CV | TimeSeriesCV(Shuffle=False, Seed=7) |
| Optimization Metric | neg_sym_mean_abs_percent_error |
| Selected Algorithm | ETSForecaster |
| Selected Hyperparameters | {'error': 'add', 'trend': 'mul', 'damped_trend': True, 'seasonal': 'add', 'seasonal_periods': 52, 'initialization_method': 'estimated', 'initial_level': None, 'initial_trend': None, 'initial_seasonal': None, 'bounds': None, 'dates': None, 'freq': None, 'missing': 'none', 'start_params': None, 'maxiter': 1500, 'disp': -1, 'return_params': False} |
| AutoMLx version | 24.3.0 |
| Python version | 3.11.9 (main, Jul 18 2024, 13:17:53) [GCC 9.3.1 20200408 (Red Hat 9.3.1-2)] |
| Step | # Samples | # Features | Algorithm | Hyperparameters | Score (neg_sym_mean_abs_percent_error) | Runtime (Seconds) | Memory Usage (GB) | Finished | |
|---|---|---|---|---|---|---|---|---|---|
| 0 | Model Selection | 161 | 2 | InputTargetDataTransformer_ETSForecaster | {'error': 'add', 'trend': 'add', 'damped_trend': False, 'seasonal': 'add', 'seasonal_periods': 52, 'initialization_method': 'estimated', 'initial_level': None, 'initial_trend': None, 'initial_seasonal': None, 'bounds': None, 'dates': None, 'freq': None, 'missing': 'none', 'start_params': None, 'maxiter': 1500, 'disp': -1, 'return_params': False} | -0.036 | 21.8211 | 0.2593 | Thu Aug 15 19:16:32 2024 |
| 1 | Model Selection | 161 | 2 | InputTargetDataTransformer_STLwARIMAForecaster | {'seasonal_deg': 1, 'trend_deg': 1, 'low_pass_deg': 0, 'period': 52, 'arima_p': 2, 'arima_d': 1, 'arima_q': 2, 'arima_trend': 'n', 'concentrate_scale': True} | -0.0855 | 3.0864 | 0.3030 | Thu Aug 15 19:16:29 2024 |
| 2 | Model Selection | 161 | 2 | InputTargetDataTransformer_XGBForecaster | {'differencing_order': 1, 'acf_local_maxima': '[52, 35, 17, 26, 22, 4, 48, 30, 13, 56, 39, 61, 44, 9, 65, 1]', 'use_X': True, 'n_estimators': 50, 'max_depth': 5} | -0.0951 | 20.1390 | 0.7001 | Thu Aug 15 19:16:34 2024 |
| 3 | Model Selection | 161 | 2 | InputTargetDataTransformer_ExtraTreesForecaster | {'differencing_order': 1, 'acf_local_maxima': '[52, 35, 17, 26, 22, 4, 48, 30, 13, 56, 39, 61, 44, 9, 65, 1]', 'use_X': True, 'n_estimators': 40, 'min_samples_leaf': 0.030458} | -0.1252 | 20.1737 | 0.6971 | Thu Aug 15 19:16:34 2024 |
| 4 | Model Selection | 161 | 2 | InputTargetDataTransformer_LGBMForecaster | {'differencing_order': 1, 'acf_local_maxima': '[52, 35, 17, 26, 22, 4, 48, 30, 13, 56, 39, 61, 44, 9, 65, 1]', 'use_X': True, 'max_depth': 4, 'n_estimators': 37} | -0.13 | 18.2716 | 0.7028 | Thu Aug 15 19:16:36 2024 |
| 5 | Model Selection | 122 | 2 | InputTargetDataTransformer_ThetaForecaster | {'sp': 52, 'deseasonalize': False, 'initial_level': None} | -0.2091 | 7.0557 | 0.2933 | Thu Aug 15 19:16:29 2024 |
| 6 | Model Selection | 148 | 2 | InputTargetDataTransformer_ExpSmoothForecaster | {'trend': 'add', 'damped_trend': True, 'seasonal': None, 'sp': 52, 'use_boxcox': False} | -0.2475 | 6.5339 | 0.2853 | Thu Aug 15 19:16:29 2024 |
| 7 | Model Selection | 109 | 2 | InputTargetDataTransformer_NaiveForecaster | {'strategy': 'last', 'sp': 52, 'window_length': None} | -0.4494 | 6.6660 | 0.2788 | Thu Aug 15 19:16:28 2024 |
| 8 | Model Selection | 161 | 2 | InputTargetDataTransformer_STLwESForecaster | {'seasonal_deg': 1, 'trend_deg': 1, 'low_pass_deg': 1, 'period': 52, 'es_trend': 'add', 'es_damped_trend': True, 'concentrate_scale': True} | -inf | 9.0492 | 0.2923 | Thu Aug 15 19:16:30 2024 |
| 9 | Model Selection | 161 | 2 | InputTargetDataTransformer_SARIMAXForecaster | {'sp': 52, 'p': 2, 'd': 1, 'q': 2, 'P': 1, 'D': 0, 'Q': 1, 'trend': 'n', 'use_X': True, 'enforce_stationarity': True, 'enforce_invertibility': True, 'method': 'lbfgs', 'disp': -1, 'concentrate_scale': False} | -inf | 32.5628 | 0.3386 | Thu Aug 15 19:16:35 2024 |
| 10 | Model Tuning | 161 | 2 | InputTargetDataTransformer_ETSForecaster | {'error': 'add', 'trend': 'mul', 'damped_trend': True, 'seasonal': 'add', 'seasonal_periods': 52, 'initialization_method': 'estimated', 'initial_level': None, 'initial_trend': None, 'initial_seasonal': None, 'bounds': None, 'dates': None, 'freq': None, 'missing': 'none', 'start_params': None, 'maxiter': 1500, 'disp': -1, 'return_params': False} | -0.0327 | 19.0839 | 0.7035 | Thu Aug 15 19:16:51 2024 |
| 11 | Model Tuning | 161 | 2 | InputTargetDataTransformer_ETSForecaster | {'error': 'add', 'trend': 'add', 'damped_trend': True, 'seasonal': 'add', 'seasonal_periods': 52, 'initialization_method': 'estimated', 'initial_level': None, 'initial_trend': None, 'initial_seasonal': None, 'bounds': None, 'dates': None, 'freq': None, 'missing': 'none', 'start_params': None, 'maxiter': 1500, 'disp': -1, 'return_params': False} | -0.0334 | 14.7964 | 0.7031 | Thu Aug 15 19:16:40 2024 |
| 12 | Model Tuning | 161 | 2 | InputTargetDataTransformer_ETSForecaster | {'error': 'mul', 'trend': 'add', 'damped_trend': True, 'seasonal': 'add', 'seasonal_periods': 52, 'initialization_method': 'estimated', 'initial_level': None, 'initial_trend': None, 'initial_seasonal': None, 'bounds': None, 'dates': None, 'freq': None, 'missing': 'none', 'start_params': None, 'maxiter': 1500, 'disp': -1, 'return_params': False} | -0.036 | 17.6693 | 0.7036 | Thu Aug 15 19:16:49 2024 |
| 13 | Model Tuning | 161 | 2 | InputTargetDataTransformer_ETSForecaster | {'error': 'add', 'trend': 'add', 'damped_trend': False, 'seasonal': 'add', 'seasonal_periods': 52, 'initialization_method': 'estimated', 'initial_level': None, 'initial_trend': None, 'initial_seasonal': None, 'bounds': None, 'dates': None, 'freq': None, 'missing': 'none', 'start_params': None, 'maxiter': 1500, 'disp': -1, 'return_params': False} | -0.0361 | 13.3357 | 0.7024 | Thu Aug 15 19:16:40 2024 |
| 14 | Model Tuning | 135 | 2 | InputTargetDataTransformer_ETSForecaster | {'error': 'add', 'trend': 'add', 'damped_trend': False, 'seasonal': 'add', 'seasonal_periods': 52, 'initialization_method': 'estimated', 'initial_level': None, 'initial_trend': None, 'initial_seasonal': None, 'bounds': None, 'dates': None, 'freq': None, 'missing': 'none', 'start_params': None, 'maxiter': 1500, 'disp': -1, 'return_params': False} | -0.0361 | 12.6781 | 0.7034 | Thu Aug 15 19:16:43 2024 |
| 15 | Model Tuning | 148 | 2 | InputTargetDataTransformer_ETSForecaster | {'error': 'add', 'trend': 'add', 'damped_trend': False, 'seasonal': 'add', 'seasonal_periods': 52, 'initialization_method': 'estimated', 'initial_level': None, 'initial_trend': None, 'initial_seasonal': None, 'bounds': None, 'dates': None, 'freq': None, 'missing': 'none', 'start_params': None, 'maxiter': 1500, 'disp': -1, 'return_params': False} | -0.0361 | 13.1573 | 0.7020 | Thu Aug 15 19:16:44 2024 |
| 16 | Model Tuning | 161 | 2 | InputTargetDataTransformer_ETSForecaster | {'error': 'add', 'trend': None, 'damped_trend': True, 'seasonal': 'add', 'seasonal_periods': 52, 'initialization_method': 'estimated', 'initial_level': None, 'initial_trend': None, 'initial_seasonal': None, 'bounds': None, 'dates': None, 'freq': None, 'missing': 'none', 'start_params': None, 'maxiter': 1500, 'disp': -1, 'return_params': False} | -0.0469 | 12.3479 | 0.7044 | Thu Aug 15 19:16:53 2024 |
| 17 | Model Tuning | 161 | 2 | InputTargetDataTransformer_ETSForecaster | {'error': 'add', 'trend': None, 'damped_trend': False, 'seasonal': 'add', 'seasonal_periods': 52, 'initialization_method': 'estimated', 'initial_level': None, 'initial_trend': None, 'initial_seasonal': None, 'bounds': None, 'dates': None, 'freq': None, 'missing': 'none', 'start_params': None, 'maxiter': 1500, 'disp': -1, 'return_params': False} | -0.0469 | 13.3312 | 0.7043 | Thu Aug 15 19:16:47 2024 |
| 18 | Model Tuning | 161 | 2 | InputTargetDataTransformer_STLwARIMAForecaster | {'seasonal_deg': 0, 'trend_deg': 1, 'low_pass_deg': 0, 'period': 52, 'arima_p': 2, 'arima_d': 1, 'arima_q': 2, 'arima_trend': 'n', 'concentrate_scale': True} | -0.0551 | 2.8733 | 0.7115 | Thu Aug 15 19:16:57 2024 |
| 19 | Model Tuning | 161 | 2 | InputTargetDataTransformer_STLwARIMAForecaster | {'seasonal_deg': 1, 'trend_deg': 1, 'low_pass_deg': 1, 'period': 52, 'arima_p': 2, 'arima_d': 1, 'arima_q': 2, 'arima_trend': 'n', 'concentrate_scale': True} | -0.0839 | 2.3810 | 0.7116 | Thu Aug 15 19:16:56 2024 |
| 20 | Model Tuning | 161 | 2 | InputTargetDataTransformer_STLwARIMAForecaster | {'seasonal_deg': 1, 'trend_deg': 1, 'low_pass_deg': 0, 'period': 52, 'arima_p': 2, 'arima_d': 1, 'arima_q': 2, 'arima_trend': 'n', 'concentrate_scale': True} | -0.0855 | 2.6034 | 0.7059 | Thu Aug 15 19:16:56 2024 |
| 21 | Model Tuning | 135 | 2 | InputTargetDataTransformer_STLwARIMAForecaster | {'seasonal_deg': 1, 'trend_deg': 1, 'low_pass_deg': 0, 'period': 52, 'arima_p': 2, 'arima_d': 1, 'arima_q': 2, 'arima_trend': 'n', 'concentrate_scale': True} | -0.0855 | 2.5809 | 0.7120 | Thu Aug 15 19:16:56 2024 |
| 22 | Model Tuning | 161 | 2 | InputTargetDataTransformer_STLwARIMAForecaster | {'seasonal_deg': 1, 'trend_deg': 1, 'low_pass_deg': 0, 'period': 52, 'arima_p': 2, 'arima_d': 1, 'arima_q': 2, 'arima_trend': 'n', 'concentrate_scale': True} | -0.0855 | 2.2540 | 0.7117 | Thu Aug 15 19:16:57 2024 |
| 23 | Model Tuning | 161 | 2 | InputTargetDataTransformer_STLwARIMAForecaster | {'seasonal_deg': 1, 'trend_deg': 1, 'low_pass_deg': 0, 'period': 52, 'arima_p': 2, 'arima_d': 1, 'arima_q': 2, 'arima_trend': 'n', 'concentrate_scale': True} | -0.0855 | 2.3840 | 0.7057 | Thu Aug 15 19:16:57 2024 |
| 24 | Model Tuning | 161 | 2 | InputTargetDataTransformer_STLwARIMAForecaster | {'seasonal_deg': 1, 'trend_deg': 1, 'low_pass_deg': 0, 'period': 52, 'arima_p': 2, 'arima_d': 1, 'arima_q': 2, 'arima_trend': 'n', 'concentrate_scale': True} | -0.0855 | 2.4665 | 0.7135 | Thu Aug 15 19:16:56 2024 |
| 25 | Model Tuning | 161 | 2 | InputTargetDataTransformer_STLwARIMAForecaster | {'seasonal_deg': 1, 'trend_deg': 1, 'low_pass_deg': 0, 'period': 52, 'arima_p': 2, 'arima_d': 1, 'arima_q': 2, 'arima_trend': 'n', 'concentrate_scale': True} | -0.0855 | 2.9517 | 0.7114 | Thu Aug 15 19:16:55 2024 |
| 26 | Model Tuning | 161 | 2 | InputTargetDataTransformer_STLwARIMAForecaster | {'seasonal_deg': 1, 'trend_deg': 1, 'low_pass_deg': 0, 'period': 52, 'arima_p': 1, 'arima_d': 1, 'arima_q': 2, 'arima_trend': 'n', 'concentrate_scale': True} | -0.0891 | 2.6177 | 0.7112 | Thu Aug 15 19:16:55 2024 |
| 27 | Model Tuning | 161 | 2 | InputTargetDataTransformer_STLwARIMAForecaster | {'seasonal_deg': 1, 'trend_deg': 1, 'low_pass_deg': 0, 'period': 52, 'arima_p': 2, 'arima_d': 1, 'arima_q': 1, 'arima_trend': 'n', 'concentrate_scale': True} | -0.0962 | 2.0384 | 0.7111 | Thu Aug 15 19:16:56 2024 |
| 28 | Model Tuning | 161 | 2 | InputTargetDataTransformer_STLwARIMAForecaster | {'seasonal_deg': 1, 'trend_deg': 1, 'low_pass_deg': 0, 'period': 52, 'arima_p': 2, 'arima_d': 1, 'arima_q': 0, 'arima_trend': 'n', 'concentrate_scale': True} | -0.1188 | 1.6989 | 0.7112 | Thu Aug 15 19:16:55 2024 |
| 29 | Model Tuning | 161 | 2 | InputTargetDataTransformer_STLwARIMAForecaster | {'seasonal_deg': 1, 'trend_deg': 1, 'low_pass_deg': 0, 'period': 52, 'arima_p': 0, 'arima_d': 1, 'arima_q': 2, 'arima_trend': 'n', 'concentrate_scale': True} | -0.1198 | 2.0501 | 0.7121 | Thu Aug 15 19:16:54 2024 |
| 30 | Model Tuning | 161 | 2 | InputTargetDataTransformer_ETSForecaster | {'error': 'add', 'trend': 'add', 'damped_trend': False, 'seasonal': 'mul', 'seasonal_periods': 52, 'initialization_method': 'estimated', 'initial_level': None, 'initial_trend': None, 'initial_seasonal': None, 'bounds': None, 'dates': None, 'freq': None, 'missing': 'none', 'start_params': None, 'maxiter': 1500, 'disp': -1, 'return_params': False} | -0.1244 | 15.2929 | 0.7039 | Thu Aug 15 19:16:44 2024 |
| 31 | Model Tuning | 161 | 2 | InputTargetDataTransformer_STLwARIMAForecaster | {'seasonal_deg': 0, 'trend_deg': 0, 'low_pass_deg': 0, 'period': 52, 'arima_p': 5, 'arima_d': 1, 'arima_q': 0, 'arima_trend': 'n', 'concentrate_scale': True} | -0.1274 | 2.2509 | 0.7121 | Thu Aug 15 19:16:58 2024 |
| 32 | Model Tuning | 161 | 2 | InputTargetDataTransformer_ETSForecaster | {'error': 'add', 'trend': 'add', 'damped_trend': True, 'seasonal': 'mul', 'seasonal_periods': 52, 'initialization_method': 'estimated', 'initial_level': None, 'initial_trend': None, 'initial_seasonal': None, 'bounds': None, 'dates': None, 'freq': None, 'missing': 'none', 'start_params': None, 'maxiter': 1500, 'disp': -1, 'return_params': False} | -0.129 | 14.4294 | 0.7005 | Thu Aug 15 19:16:49 2024 |
| 33 | Model Tuning | 161 | 2 | InputTargetDataTransformer_STLwARIMAForecaster | {'seasonal_deg': 0, 'trend_deg': 0, 'low_pass_deg': 0, 'period': 52, 'arima_p': 1, 'arima_d': 1, 'arima_q': 0, 'arima_trend': 'n', 'concentrate_scale': True} | -0.135 | 1.6667 | 0.7119 | Thu Aug 15 19:16:58 2024 |
| 34 | Model Tuning | 161 | 2 | InputTargetDataTransformer_STLwARIMAForecaster | {'seasonal_deg': 0, 'trend_deg': 0, 'low_pass_deg': 0, 'period': 52, 'arima_p': 2, 'arima_d': 1, 'arima_q': 0, 'arima_trend': 'n', 'concentrate_scale': True} | -0.1359 | 1.8356 | 0.7061 | Thu Aug 15 19:16:58 2024 |
| 35 | Model Tuning | 161 | 2 | InputTargetDataTransformer_STLwARIMAForecaster | {'seasonal_deg': 0, 'trend_deg': 0, 'low_pass_deg': 0, 'period': 52, 'arima_p': 3, 'arima_d': 1, 'arima_q': 0, 'arima_trend': 'n', 'concentrate_scale': True} | -0.1376 | 2.4579 | 0.7141 | Thu Aug 15 19:16:58 2024 |
| 36 | Model Tuning | 161 | 2 | InputTargetDataTransformer_ETSForecaster | {'error': 'add', 'trend': 'mul', 'damped_trend': False, 'seasonal': 'add', 'seasonal_periods': 52, 'initialization_method': 'estimated', 'initial_level': None, 'initial_trend': None, 'initial_seasonal': None, 'bounds': None, 'dates': None, 'freq': None, 'missing': 'none', 'start_params': None, 'maxiter': 1500, 'disp': -1, 'return_params': False} | -0.1409 | 15.7406 | 0.7036 | Thu Aug 15 19:16:47 2024 |
| 37 | Model Tuning | 161 | 2 | InputTargetDataTransformer_ETSForecaster | {'error': 'add', 'trend': 'add', 'damped_trend': True, 'seasonal': 'add', 'seasonal_periods': 1, 'initialization_method': 'estimated', 'initial_level': None, 'initial_trend': None, 'initial_seasonal': None, 'bounds': None, 'dates': None, 'freq': None, 'missing': 'none', 'start_params': None, 'maxiter': 1500, 'disp': -1, 'return_params': False} | -0.2209 | 1.4717 | 0.7033 | Thu Aug 15 19:16:48 2024 |
| 38 | Model Tuning | 161 | 2 | InputTargetDataTransformer_ETSForecaster | {'error': 'add', 'trend': 'add', 'damped_trend': True, 'seasonal': None, 'seasonal_periods': 52, 'initialization_method': 'estimated', 'initial_level': None, 'initial_trend': None, 'initial_seasonal': None, 'bounds': None, 'dates': None, 'freq': None, 'missing': 'none', 'start_params': None, 'maxiter': 1500, 'disp': -1, 'return_params': False} | -0.2209 | 1.4841 | 0.7041 | Thu Aug 15 19:16:47 2024 |
| 39 | Model Tuning | 161 | 2 | InputTargetDataTransformer_STLwARIMAForecaster | {'seasonal_deg': 1, 'trend_deg': 1, 'low_pass_deg': 0, 'period': 35, 'arima_p': 2, 'arima_d': 1, 'arima_q': 2, 'arima_trend': 'n', 'concentrate_scale': True} | -0.2295 | 2.0397 | 0.7116 | Thu Aug 15 19:16:56 2024 |
| 40 | Model Tuning | 161 | 2 | InputTargetDataTransformer_STLwARIMAForecaster | {'seasonal_deg': 1, 'trend_deg': 0, 'low_pass_deg': 0, 'period': 52, 'arima_p': 2, 'arima_d': 1, 'arima_q': 2, 'arima_trend': 'n', 'concentrate_scale': True} | -0.2544 | 2.4226 | 0.7123 | Thu Aug 15 19:16:57 2024 |
| 41 | Model Tuning | 161 | 2 | InputTargetDataTransformer_ETSForecaster | {'error': 'add', 'trend': 'add', 'damped_trend': False, 'seasonal': 'add', 'seasonal_periods': 1, 'initialization_method': 'estimated', 'initial_level': None, 'initial_trend': None, 'initial_seasonal': None, 'bounds': None, 'dates': None, 'freq': None, 'missing': 'none', 'start_params': None, 'maxiter': 1500, 'disp': -1, 'return_params': False} | -0.2919 | 1.5223 | 0.7032 | Thu Aug 15 19:16:43 2024 |
| 42 | Model Tuning | 161 | 2 | InputTargetDataTransformer_ETSForecaster | {'error': 'add', 'trend': 'add', 'damped_trend': False, 'seasonal': None, 'seasonal_periods': 52, 'initialization_method': 'estimated', 'initial_level': None, 'initial_trend': None, 'initial_seasonal': None, 'bounds': None, 'dates': None, 'freq': None, 'missing': 'none', 'start_params': None, 'maxiter': 1500, 'disp': -1, 'return_params': False} | -0.2919 | 1.4820 | 0.7017 | Thu Aug 15 19:16:41 2024 |
| 43 | Model Tuning | 161 | 2 | InputTargetDataTransformer_ETSForecaster | {'error': 'add', 'trend': 'add', 'damped_trend': False, 'seasonal': 'add', 'seasonal_periods': 35, 'initialization_method': 'estimated', 'initial_level': None, 'initial_trend': None, 'initial_seasonal': None, 'bounds': None, 'dates': None, 'freq': None, 'missing': 'none', 'start_params': None, 'maxiter': 1500, 'disp': -1, 'return_params': False} | -inf | 14.4222 | 0.7011 | Thu Aug 15 19:16:45 2024 |
| 44 | Model Tuning | 161 | 2 | InputTargetDataTransformer_ETSForecaster | {'error': 'add', 'trend': 'add', 'damped_trend': False, 'seasonal': 'add', 'seasonal_periods': 52, 'initialization_method': 'estimated', 'initial_level': None, 'initial_trend': None, 'initial_seasonal': None, 'bounds': None, 'dates': None, 'freq': None, 'missing': 'none', 'start_params': None, 'maxiter': 1500, 'disp': -1, 'return_params': False} | -inf | 14.5241 | 0.7018 | Thu Aug 15 19:16:47 2024 |
| 45 | Model Tuning | 161 | 2 | InputTargetDataTransformer_ETSForecaster | {'error': 'add', 'trend': 'add', 'damped_trend': False, 'seasonal': 'add', 'seasonal_periods': 52, 'initialization_method': 'estimated', 'initial_level': None, 'initial_trend': None, 'initial_seasonal': None, 'bounds': None, 'dates': None, 'freq': None, 'missing': 'none', 'start_params': None, 'maxiter': 1500, 'disp': -1, 'return_params': False} | -inf | 16.0069 | 0.7010 | Thu Aug 15 19:16:41 2024 |
| 46 | Model Tuning | 161 | 2 | InputTargetDataTransformer_ETSForecaster | {'error': 'add', 'trend': 'add', 'damped_trend': True, 'seasonal': 'add', 'seasonal_periods': 35, 'initialization_method': 'estimated', 'initial_level': None, 'initial_trend': None, 'initial_seasonal': None, 'bounds': None, 'dates': None, 'freq': None, 'missing': 'none', 'start_params': None, 'maxiter': 1500, 'disp': -1, 'return_params': False} | -inf | 14.5421 | 0.7042 | Thu Aug 15 19:16:50 2024 |
| 47 | Model Tuning | 161 | 2 | InputTargetDataTransformer_ETSForecaster | {'error': 'mul', 'trend': 'add', 'damped_trend': False, 'seasonal': 'add', 'seasonal_periods': 52, 'initialization_method': 'estimated', 'initial_level': None, 'initial_trend': None, 'initial_seasonal': None, 'bounds': None, 'dates': None, 'freq': None, 'missing': 'none', 'start_params': None, 'maxiter': 1500, 'disp': -1, 'return_params': False} | -inf | 22.2689 | 0.7002 | Thu Aug 15 19:16:42 2024 |
| 48 | Model Tuning | 161 | 2 | InputTargetDataTransformer_STLwARIMAForecaster | {'seasonal_deg': 0, 'trend_deg': 0, 'low_pass_deg': 0, 'period': 52, 'arima_p': 0, 'arima_d': 1, 'arima_q': 0, 'arima_trend': 'n', 'concentrate_scale': True} | -inf | 1.4146 | 0.7141 | Thu Aug 15 19:16:57 2024 |
| 49 | Model Tuning | 161 | 2 | InputTargetDataTransformer_STLwARIMAForecaster | {'seasonal_deg': 0, 'trend_deg': 0, 'low_pass_deg': 0, 'period': 52, 'arima_p': 0, 'arima_d': 1, 'arima_q': 0, 'arima_trend': 'n', 'concentrate_scale': True} | -inf | 1.4281 | 0.7120 | Thu Aug 15 19:16:59 2024 |
| 50 | Model Tuning | 161 | 2 | InputTargetDataTransformer_STLwARIMAForecaster | {'seasonal_deg': 1, 'trend_deg': 1, 'low_pass_deg': 0, 'period': 1, 'arima_p': 2, 'arima_d': 1, 'arima_q': 2, 'arima_trend': 'n', 'concentrate_scale': True} | -inf | 1.5674 | 0.7139 | Thu Aug 15 19:16:56 2024 |
| 51 | Model Tuning | 161 | 2 | InputTargetDataTransformer_STLwARIMAForecaster | {'seasonal_deg': 1, 'trend_deg': 1, 'low_pass_deg': 0, 'period': 52, 'arima_p': 3, 'arima_d': 1, 'arima_q': 2, 'arima_trend': 'n', 'concentrate_scale': True} | -inf | 3.3569 | 0.7005 | Thu Aug 15 19:16:55 2024 |
| 52 | Model Tuning | 122 | 2 | InputTargetDataTransformer_STLwARIMAForecaster | {'seasonal_deg': 1, 'trend_deg': 1, 'low_pass_deg': 0, 'period': 52, 'arima_p': 5, 'arima_d': 1, 'arima_q': 2, 'arima_trend': 'n', 'concentrate_scale': True} | -inf | 3.7713 | 0.7135 | Thu Aug 15 19:16:56 2024 |
We also provide the capability to visualize the results of each stage of the AutoML pipeline.
The plot below shows the scores predicted by Algorithm Selection for each algorithm. Since negative sMAPE is used by default, higher values (closer to zero) are better. The horizontal line shows the average score across all algorithms. Algorithms with better score than average are colored turquoise, whereas those with worse score than average are colored teal.
# Each trial is a tuple of
# (algorithm, no. samples, no. features, mean CV score, hyperparameters,
# all CV scores, total CV time (s), memory usage (Gb))
trials = est1.completed_trials_summary_[est1.completed_trials_summary_["Step"].str.contains('Model Selection')]
name_of_score_column = f"Score ({est1._inferred_score_metric[0].name})"
trials.replace([np.inf, -np.inf], np.nan, inplace=True)
trials.dropna(subset=[name_of_score_column], inplace = True)
scores = trials[name_of_score_column].tolist()
models = trials['Algorithm'].tolist()
colors = []
y_margin = 0.10 * (max(scores) - min(scores))
s = pd.Series(scores, index=models).sort_values(ascending=False)
s = s.dropna()
for f in s.keys():
if f.strip() == est1.selected_model_.strip():
colors.append('orange')
elif s[f] >= s.mean():
colors.append('teal')
else:
colors.append('turquoise')
fig, ax = plt.subplots(1)
ax.set_title("Algorithm Selection Trials")
ax.set_ylim(min(scores) - y_margin, max(scores) + y_margin)
ax.set_ylabel(est1._inferred_score_metric[0].name)
s.plot.bar(ax=ax, color=colors, edgecolor='black')
ax.axhline(y=s.mean(), color='black', linewidth=0.5)
plt.show()
Hyperparameter tuning is the last stage of the Oracle AutoMLx pipeline, and focuses on improving the chosen algorithm's score. We use a novel algorithm to search across many hyperparameter dimensions, and converge automatically when optimal hyperparameters are identified. Each trial in the graph below represents a particular hyperparameter combination for the selected model.
# Each trial is a row in a dataframe that contains
# Algorithm, Number of Samples, Number of Features, Hyperparameters, Score, Runtime, Memory Usage, Step as features
trials = est1.completed_trials_summary_[est1.completed_trials_summary_["Step"].str.contains('Model Tuning')]
trials.replace([np.inf, -np.inf], np.nan, inplace=True)
trials.dropna(subset=[name_of_score_column], inplace = True)
trials.drop(trials[trials['Finished'] == -1].index, inplace = True)
trials['Finished']= trials['Finished'].apply(lambda x: time.mktime(datetime.datetime.strptime(x,
"%a %b %d %H:%M:%S %Y").timetuple()))
trials.sort_values(by=['Finished'],ascending=True, inplace = True)
scores = trials[name_of_score_column].tolist()
score = []
score.append(scores[0])
for i in range(1,len(scores)):
if scores[i]>= score[i-1]:
score.append(scores[i])
else:
score.append(score[i-1])
y_margin = 0.10 * (max(score) - min(score))
fig, ax = plt.subplots(1)
ax.set_title("Hyperparameter Tuning Trials")
ax.set_xlabel("Iteration $n$")
ax.set_ylabel(est1._inferred_score_metric[0].name)
ax.grid(color='g', linestyle='-', linewidth=0.1)
ax.set_ylim(min(score) - y_margin, max(score) + y_margin)
ax.plot(range(1, len(trials) + 1), score, 'k:', marker="s", color='teal', markersize=3)
plt.show()
We can also view all tuned algorithms, as well as their validation and testing performance. This provides a good sanity check for the decision making.
print(f'### Plotting is enabled for total models tuned = {len(est1.pipelines_)}.')
print('### Model_name\t\t Val_score ')
model_name = ""
if len(est1.pipelines_) > 1:
model_name = "InputTargetDataTransformer_"
for pipeline in est1.pipelines_:
print( pipeline.selected_model_, " \t",
'%.4f'%est1.completed_trials_summary_[est1.completed_trials_summary_["Algorithm"] == model_name+pipeline.selected_model_].iloc[0][name_of_score_column],'\t', # validation_score
)
predictions = pipeline.forecast(periods=len(y_test), alpha=0.05) # out-of-sample forecast
predictions.index = y_test.index
fig = pipeline.plot_forecast(predictions=predictions, additional_frames=dict(y_test=y_test))
fig.show()
### Plotting is enabled for total models tuned = 2. ### Model_name Val_score ETSForecaster -0.0360
STLwARIMAForecaster -0.0855
We now display the use of the Oracle AutoMLx solution for multivariate time series.
There are two types of multivariate time series in forecasting:
A time series target is the variable that the forecaster will try to predict. Additional variables provide more information that help with the prediction of the target. In single-target forecasting, we are forecasting a single variable, whereas in multi-target forecasting, we forecast values of highly correlated variables together as a system.
We will first see an example of single-target forecasting with additional variables and then a second example with a multi-target forecasting problem.
We will now take two time series from the M4 dataset: one target variable, and one additional variable used solely as explanatory variables. Note that $X$ consists of the additional variables ('W142') while $y$ has the target variables ('W143'). We will also showcase AutoML's functionality in the absence of a datetime index, by dropping the datetime index and utilize the series with only an int64index.
series_ids = ["W142", "W143"]
series_values = all_series.loc[series_ids].T
# drop NaNs for the time period where data wasn't recorded
series_values.dropna(inplace=True)
# we will take the first 100 elements of the series
series_values = series_values[:100]
# retrieve starting date of recording and series length to generate the datetime index
start_date = "28-06-00 12:00"
future_dates = pd.date_range(start=start_date, periods=len(series_values), freq="W")
data = pd.DataFrame(series_values.to_numpy(), index=future_dates, columns=series_ids)
data.index = np.arange(0, len(data))
y = pd.DataFrame(data["W142"])
X = pd.DataFrame(data["W143"])
We then split it using a temporal train-test split as done previously.
X_train_df, X_test_df = temporal_train_test_split(X, train_size=0.9)
y_train_df, y_test_df = temporal_train_test_split(y, train_size=0.9)
We will now fit the pipeline on this data. For the single-target forecasting task with additional variables, the pipeline considers all available models by default. However, the models that actually support additional variables are the following:
LGBMForecasterExtraTreesForecasterXGBForecasterSARIMAXForecasterProphetForecaster
These models also have the option to discard additional variables. This can be set as a boolean hyperparameter called use_X. est2 = automlx.Pipeline(task='forecasting',
model_list=['LGBMForecaster',
'ExtraTreesForecaster',
'XGBForecaster',
'SARIMAXForecaster',
'ProphetForecaster'])
est2.fit(X=X_train_df, y=y_train_df)
print('Selected model: {}'.format(est2.selected_model_))
print('Selected model params: {}'.format(est2.selected_model_params_))
[2024-08-15 19:17:12,297] [automlx.data_transform] Number of simple differencing orders required: d = 0
[2024-08-15 19:17:12,300] [automlx.data_transform] Seasonal Periodicities; from decomposed/adjusted: [1]
[2024-08-15 19:17:12,549] [automlx.interface] Dataset shape: (90,3)
[2024-08-15 19:17:12,877] [automlx.data_transform] Running preprocessing. Number of features: 4
[2024-08-15 19:17:13,134] [automlx.data_transform] Preprocessing completed. Took 0.257 secs
[2024-08-15 19:17:13,439] [automlx.process] Running Model Generation
[2024-08-15 19:17:13,520] [automlx] Provided model (SARIMAXForecaster) is not supported.Supported models are: ['LGBMForecaster', 'ExtraTreesForecaster', 'ExpSmoothForecaster', 'NaiveForecaster', 'ThetaForecaster', 'ETSForecaster', 'XGBForecaster']
[2024-08-15 19:17:13,521] [automlx] Provided model (ProphetForecaster) is not supported.Supported models are: ['LGBMForecaster', 'ExtraTreesForecaster', 'ExpSmoothForecaster', 'NaiveForecaster', 'ThetaForecaster', 'ETSForecaster', 'XGBForecaster']
[2024-08-15 19:17:13,523] [automlx.process] Model Generation completed.
[2024-08-15 19:17:13,689] [automlx.model_selection] Running Model Selection
[2024-08-15 19:17:21,086] [automlx.model_selection] Model Selection completed - Took 7.398 sec - Selected models: ['XGBForecaster']
[2024-08-15 19:17:21,567] [automlx.trials] Running Model Tuning for ['XGBForecaster']
[2024-08-15 19:17:36,683] [automlx.trials] Best parameters for XGBForecaster: {'differencing_order': 1, 'acf_local_maxima': '[4, 9, 18, 22, 14, 31, 26, 12, 1]', 'use_X': True, 'n_estimators': 16, 'max_depth': 2}
[2024-08-15 19:17:36,685] [automlx.trials] Model Tuning completed. Took: 15.118 secs
[2024-08-15 19:17:37,868] [automlx.interface] Re-fitting pipeline
[2024-08-15 19:17:41,159] [automlx.interface] AutoMLx completed.
Selected model: XGBForecaster
Selected model params: {'differencing_order': 1, 'acf_local_maxima': '[4, 9, 18, 22, 14, 31, 26, 12, 1]', 'use_X': True, 'n_estimators': 16, 'max_depth': 2}
For this use case, we now take 3 variables from the M4 dataset: two target variables (W142, W143) and one additional variables (W141). We will add a constant variable as an additional variable, so we have a total of two additional variables. We then split the data using a temporal train-test split and drop the datetime index, as done previously,
series_ids = ["W141", "W142", "W143"]
series_values = all_series.loc[series_ids].T
# drop NaNs for the time period where data wasn't recorded
series_values.dropna(inplace=True)
# we will take the first 100 elements of the series
series_values = series_values[:100]
# retrieve starting date of recording and series length to generate the datetime index
start_date = "28-06-00 12:00"
future_dates = pd.date_range(start=start_date, periods=len(series_values), freq="W")
data = pd.DataFrame(series_values.to_numpy(), index=future_dates, columns=series_ids)
data.index = np.arange(0, len(data))
y = data[["W142", "W143"]]
X = pd.DataFrame(data["W141"])
# adding the constant variable
X["const"] = 1
X_train_df, X_test_df = temporal_train_test_split(X, train_size=0.9)
y_train_df, y_test_df = temporal_train_test_split(y, train_size=0.9)
You can also configure the pipeline with suitable parameters according to your needs.
For the Multi-target forecasting task, the pipeline only considers two models : VARMAX and DynFactor.
custom_pipeline = automlx.Pipeline(
task='forecasting', # Supports 'classification', regression', 'anomaly_detection' and 'forecasting'
model_list=[ # Specify the models you want the AutoMLx to consider
'DynFactorForecaster',
'VARMAXForecaster'
],
n_algos_tuned=2, # Choose how many models to tune
preprocessing=True, # Disable or enable Preprocessing step. Default to `True`
search_space={}, # You can specify the hyper-parameters and ranges we search
max_tuning_trials=2, # The maximum number of tuning trials. Can be integer or Dict (max number for each model)
score_metric='neg_sym_mean_abs_percent_error', # Any scikit-learn metric or a custom function
time_series_period=None # The period of time series
)
AutoML automatically decides how many folds to create, given the length of the input series. This is dependent on the frequency and length of the series. In the above, the preprocessor chose to create two folds. In the following we set the number of folds to 8.
custom_pipeline.fit(X=X_train_df, y=y_train_df, cv=8)
[2024-08-15 19:17:41,811] [automlx.data_transform] Number of simple differencing orders required: d = 0 [2024-08-15 19:17:41,815] [automlx.data_transform] Seasonal Periodicities; from decomposed/adjusted: [1] [2024-08-15 19:17:42,153] [automlx.interface] Dataset shape: (90,5) [2024-08-15 19:17:42,229] [automlx.interface] Model Tune disabled. [2024-08-15 19:17:42,486] [automlx.data_transform] Running preprocessing. Number of features: 6 [2024-08-15 19:17:42,713] [automlx.data_transform] Preprocessing completed. Took 0.227 secs [2024-08-15 19:17:43,029] [automlx.process] Running Model Generation [2024-08-15 19:17:43,143] [automlx.process] Model Generation completed. [2024-08-15 19:17:43,555] [automlx.model_selection] Running Model Selection [2024-08-15 19:17:43,555] [automlx.trials] Fewer models (2) than top_k (2) provided, skipping model selection [2024-08-15 19:17:43,644] [automlx.interface] Re-fitting pipeline [2024-08-15 19:17:46,091] [automlx.interface] AutoMLx completed.
<automlx._interface.forecaster.AutoForecaster at 0x7efcb15d9ed0>
The AutoML pipeline provides attributes to get the selected features, the chosen model, hyperparameters as well as the score on the test set.
test_score = custom_pipeline.score(X=X_test_df,y=y_test_df)
print('Ranked models: {}'.format(custom_pipeline.ranked_models_))
print('Selected model: {}'.format(custom_pipeline.selected_model_))
print('Selected model params: {}'.format(custom_pipeline.selected_model_params_))
print(f'Score on test data : {test_score}')
[2024-08-15 19:17:46,479] [automlx.model] Predict called on DynFactorForecaster that's not been fitted.
Ranked models: ['DynFactorForecaster', 'VARMAXForecaster']
Selected model: DynFactorForecaster
Selected model params: {'k_factors': 1, 'error_order': 2, 'factor_order': 2, 'error_cov_type': 'diagonal', 'use_X': True, 'error_var': False, 'enforce_stationarity': True}
Score on test data : -0.4736945883029529
As mention in the univariate data case, there are two ways of making a prediction :
forecast(k) allows one to predict k steps after the end of the training data. It should be used when one wants to make out-of-sample predictionspredict(X) returns predictions at the timestamps given as argument. It should be used when one wants to make in-sample predictions and out-of-sample predictions. It does not support confidence intervals.In the cell below predict() is used on the last 5 timestamps of the train set, and all timestamps of the test set. forecast() is used to predict k steps after the training set, where k is the size of the test set.
y_pred = custom_pipeline.predict(pd.concat([X_train_df[-5:0],X_test_df], axis=0) )
y_forecast = custom_pipeline.forecast(len(y_test_df), alpha=0.8, X=X_test_df) # out-of-sample forecast
[2024-08-15 19:17:46,597] [automlx.model] Predict called on DynFactorForecaster that's not been fitted. [2024-08-15 19:17:46,639] [automlx.model] Forecast called on DynFactorForecaster that's not been fitted.
The obtained forecast contains predictions for the two target variables, as well as lower and upper confidence intervals, for each timestamp in the test set.
y_forecast
| W142 | W143 | W142_ci_lower | W142_ci_upper | |
|---|---|---|---|---|
| 90 | 1523.315370 | 4120.632282 | 1514.232711 | 1532.398028 |
| 91 | 1529.220991 | 4110.701763 | 1515.520362 | 1542.921620 |
| 92 | 1516.883719 | 4118.868327 | 1499.729383 | 1534.038055 |
| 93 | 1512.790195 | 4119.127842 | 1492.815650 | 1532.764741 |
| 94 | 1512.239464 | 4115.999969 | 1489.846346 | 1534.632581 |
| 95 | 1504.476192 | 4119.899345 | 1479.947498 | 1529.004886 |
| 96 | 1504.039436 | 4116.732328 | 1477.588597 | 1530.490274 |
| 97 | 1498.492695 | 4118.553261 | 1470.288723 | 1526.696668 |
| 98 | 1502.087490 | 4111.546114 | 1472.268694 | 1531.906287 |
| 99 | 1493.464952 | 4116.418334 | 1462.147539 | 1524.782365 |
One can also directly compute the score of the tuned model on the test set, without needing to run forecast() or predict().
print("Tuned model testing score (negative sMAPE): ", custom_pipeline.score(X=X_test_df, y=y_test_df))
[2024-08-15 19:17:46,844] [automlx.model] Predict called on DynFactorForecaster that's not been fitted. Tuned model testing score (negative sMAPE): -0.4736945883029529
Finally, when given as input the forecasted variables, the plot_forecast() method displays an interactive plot of the predictions (for each target variable) and confidence intervals.
custom_pipeline.plot_forecast(predictions=y_forecast, additional_frames=dict(test=y_test_df))
For a variety of decision-making tasks, getting only a prediction as model output is not sufficient. A user may wish to know why the model outputs that prediction, or which data features are relevant for that prediction. For that purpose the Oracle AutoMLx solution defines the MLExplainer object, which allows to compute multiple types of model explanations.
The MLExplainer object takes as arguments the trained model, the training data and labels, as well as the task. Let's assume that we want to learn more about the decision-making process of the model est1 that we trained based on the M4 dataset (Finance_W142) previously. Thus, we pass est1 along with its training series (y_train) to the MLExplainer to produce the appropriate explainer.
explainer = automlx.MLExplainer(est1,
None,
y_train,
task="forecasting")
For any data point within the forecasting horizon, explanations indicate which factors increased or decreased the model's prediction compared to a previous reference point. This comparison provides additional context, making the information more interpretable for the user.
The function explain_prediction() calculates the comparative feature importance for a specified datapoint. For example, if you want to explore the model's predictions and understand the impact of various temporal concepts on the decision-making process for the forecast 10 time steps in the future, you can obtain the prediction explanation as shown below.
exp = explainer.explain_prediction(forecast_timepoints=[10])
There are two options to show the explanation's results:
to_dataframe() will return a dataframe of the results. The concepts and features are returned in decreasing order of importance.show_in_notebook() will present the results as a Waterfall plot, comparing the prediction to the last observation from the training data (the reference point). This chart highlights the impact of each feature or concept on shifting the forecast from the reference point to our target prediction. Users can track each factor's contribution, categorized as positive (in red) or negative (in gray). By following the arrows, users can see how different elements affect the model's output, aiding in informed decision-making. A higher importance score indicates a stronger link between a feature and the forecast difference. Hovering over the bars reveals more details on each factor's effect on the results.For example, starting with the known reference value, the analysis from the explainer reveals that the primary factor influencing the trained model is yearly periodicity. Yearly periodicity, which refers to the recurring patterns observed each year, has contributed to an increase in the forecast. Following this, the data trend, indicating the general direction in which the data is moving over time, also plays a significant role. The time series shows an upward trend between the reference and the prediction, leading the model to predict an increase in the forecast. Consequently, the final prediction is higher than the reference value.
exp[0].to_dataframe(n_features=10)
[2024-08-15 19:17:52,129] [automlx.mlx] Displaying only non-zero attributions.
| Feature | Attribution | |
|---|---|---|
| 0 | Yearly periodicity of Finance_W142 | 70.394186 |
| 1 | Trend of Finance_W142 | 18.605814 |
exp[0].show_in_notebook()
The above explanation provided a comparative analysis of the specified forecast time point versus the default reference time point, which is automatically set to the last observation in the training data. However, advanced users can change the reference and set it to another observation (-13, the most recent peak, in this example) in the training data as shown below.
advanced_exp = explainer.explain_prediction(forecast_timepoints=[10], reference_timepoint=-13)
advanced_exp[0].show_in_notebook()